Google’s AGI Safety and Alignment Team Built a Workaround for Its Own Hiring Filter

Google DeepMind’s AGI Safety and Alignment Team built its own workaround for a hiring filter it no longer trusted. According to an internal document viewed by Bloomberg and reported by HR Executive, members of the team advised internal candidates applying to join them to fill out a second form alongside the standard application. HR Executive reads the memo’s language as suggesting candidates could be improperly screened out, raising doubt about the dependability of Google’s AI hiring process. The document states the reason plainly: “We have an applications system with a non-trivial probability your CV will be screened out incorrectly or take too long to reach us.”

What the Memo Told Candidates To Do

The secondary form guaranteed something the main pipeline could not: a human set of eyes. “Filling out this form makes sure that a real human on the team will get to see your application,” the document reportedly said. It came with a second instruction: sound like yourself, not like a model. The same document warned candidates to keep AI use light, because “these humans get really tired of reading LLM answers, because they all sound very same-y.”

Look at what is stacked here. One AI system screens candidates. A second, human-only form exists specifically to catch that system’s mistakes. Candidates are coached to write like people rather than machines, so the human backstop judges them fairly. That is three layers of mistrust wrapped around a single hiring decision, inside one team.

A Trust Problem That Runs Both Directions

HR Executive frames this inside a wider trust breakdown, and the numbers cut both ways. Gartner finds only about a quarter of candidates believe AI tools will evaluate their application fairly. Candidates are not passive in this either: close to 40% told Gartner they have used AI during the application process, most often to build a resume. And employers are not being straight about their side, with Resume Genius finding that 87% of US hiring managers use AI in hiring while roughly one in five never tell candidates how it shapes decisions about them.

None of that is new to anyone watching the AI screening debate, including the lawsuits alleging such tools disadvantage certain job seekers. What is new is who is acting like the system cannot be relied on. Not an outside advocate. Not a plaintiff’s lawyer. The company’s own people, patching around it themselves rather than flagging it as broken.

The Candidates Nobody Counted

Here is what the coverage does not measure, and it matters more than the memo itself. If one team inside Google felt it needed a private channel to catch what its own filter was missing, how many qualified candidates on other teams, with no internal advocate to build them one, got screened out and nobody ever noticed?

A back channel like this produces zero visibility. It generates no support ticket and no data point in a quality-of-hire review. It moves a handful of known, connected candidates around a system that everyone else applying through the front door still has to trust blindly, and tells Google nothing about how often the filter is wrong, or for whom.

What Accountability Would Actually Require

In our work with hiring teams, the people closest to a screening tool are usually first to sense when it produces false negatives, often well before that shows up in a quality-of-hire number. When those people respond by building a private fix instead of escalating the fault, the problem sits with governance. It has nothing to do with how candidates behave, and everything to do with who gets told when a tool stops working.

The DeepMind team did not set out to game the system. It set out to get around a system its own memo already distrusted, one it could not interrogate either.

An AI hiring filter nobody inside the building can explain is not a tool anyone should trust, no matter how many resumes it clears in a day. That Google’s filter has bugs is not the lesson here. Every filter does. The lesson is that the fix for those bugs was a private form circulated inside one team, not an audit trail every recruiter could see. That gap means the organization has already admitted it cannot account for its own hiring decisions.

Auditable does not mean a dashboard showing how many resumes got processed. It means answering, for any single rejected candidate, exactly which signal the filter weighted, and showing that the pattern held for similar candidates too. Try asking your own screening layer for that answer and see how far you get.

Without it, recruiters will do what Google’s team did: quietly build a side door for the people they happen to know, and stay silent about everyone else. The candidates it never reaches are the ones this story cannot count, and that gap is the real cost of a screening system nobody can fully vouch for.

Original reporting: HR Executive.